Objectives A small sample intelligent fault diagnosis method for rolling bearings based on phase space reconstruction (PSR) combined with principal component analysis (PCA) and an improved convolutional neural network (CNN) is proposed to address the problems of traditional fault diagnosis methods being unable to fully explore the fault information contained in one-dimensional nonlinear temporal vibration signals, the large number of samples required for model training, and insufficient generalization.

Methods First, PSR based on chaos theory is used to recover the potential dynamic features of the data, achieving the high-dimensional mapping of signals in phase space. Next, PCA is used to reduce data dimensionality redundancy and generate concise and informative fault feature reconstruction phase diagrams. Finally, CNN is used to automatically learn and extract features from complex data, achieving the intelligent fault diagnosis of rolling bearings in small sample scenarios.

Results The PSR-PCA-CNN method is validated using two bearing datasets, with experimental diagnostic accuracy exceeding 97%. In a small sample scenario with a 10% training set, the testing accuracy is higher than 90%. Compared with other feature extraction methods and intelligent algorithms, the PSR-PCA-CNN method has higher experimental accuracy.

Conclusions Compared with other image encoding methods and intelligent algorithms, the intelligent diagnostic method proposed herein has good diagnostic performance and a superior ability to extract sample features in small sample scenarios, making it an effective model for solving the small sample fault diagnosis task of rolling bearings.